Search

At present, the use of eye-tracking data in immersive Virtual Reality (iVR) learning environments is set to become a powerful tool for maximizing learning outcomes, due to the low-intrusiveness of eye-tracking technology and its integration in commercial iVR Head Mounted Displays. However, the most suitable technologies for data processing should first be identified before their use in learning environments can be generalized. In this research, the use of machine-learning techniques is proposed for that purpose, evaluating their capabilities to classify the quality of the learning environment and to predict user learning performance. To do so, an iVR learning experience simulating the operation of a bridge crane was developed. Through this experience, the performance of 63 students was evaluated, both under optimum learning conditions and under stressful conditions. The final dataset included 25 features, mostly temporal series, with a dataset size of up to 50M data points. The results showed that different classifiers (KNN, SVM and Random Forest) provided the highest accuracy when predicting learning performance variations, while the accuracy of user learning performance was still far from optimized, opening a new line of future research. This study has the objective of serving as a baseline for future improvements to model accuracy using complex machine-learning techniques. [ABSTRACT FROM AUTHOR]

This article presents the calculation of the sdg 11.3.1 and 11.7.1 indicators in Colombia, integrating statistical and geospatial information, as essential sources to achieve a robust and territorially disaggregated measurement. Based on the processes defined by un-Habitat, it develops a methodology with a geospatial emphasis, supported by the processing of satellite images through the Random Forest supervised classification algorithm, to obtain the metrics required in the calculation of the two indicators, such as built-up areas, urban land consumption, and open spaces, together with integrated analysis of statistical information. The sdg 11.3.1 indicator for the 2015-2020 period was calculated for 63 cities, whose national value of 0.43 highlights that efficient land use is made in the country, while sdg 11.7.1 for the year 2018 was calculated in a representative sample of nine cities, which indicates that at the national level 33.2 % of built areas are allocated to open spaces for public use. These results make the country a regional benchmark in the monitoring of the sdgs, highlighting the possibility of updating the results in the future, thanks to automated processing in the cloud using scripts developments. [ABSTRACT FROM AUTHOR]

This work presents applications of the Backward Seismic Analysis (BSA) method for steel storage tanks using a data base of more than 382 steel storage tanks in operation during large subductive earthquakes: Valdivia 1960, Central Chile 1985, Tocopilla 2007, El Maule 2010, Alaska 1964, and others in the United States between 1933 and 1995 (subductive and cortical). It has been recorded that most of the steel storage tanks without anchor systems have failed during large earthquakes. These have been designed with the standards API 650-E, AWWA-D100, and NZSEE, which propose similar procedures for estimating seismic forces, but with different design methods. During different conferences, the causes of the failures were evaluated, concluding that the tanks were designed mainly with the API 650-E code and were unanchored. Moreover, the design codes employed do not consider relevant aspects that condition the seismic response of steel storage tanks. This work develops a prediction model based on the historical information already described, which is capable of efficiently predicting if a steel storage tank will suffer any failures during an earthquake. Various algorithms were evaluated, finding that the Random Forest method exhibits the best results. The results obtained in the prediction of steel storage tank failures reach more than 90% efficiency in most of the evaluated scenarios. [ABSTRACT FROM AUTHOR]

In this work we consider 148 semiochemicals reported for the family Scarabaeidae, whose chemical structure was characterized using a set of 200 molecular descriptors from five different classes. The selection of the most discriminating descriptors was carried out with three different techniques: Principal Component Analysis, for each class of descriptors, Random Forests and Boruta-Shap, applied to the total of descriptors. Although the three techniques are conceptually different, they select a similar number of descriptors from each class. We proposed a combination of machine learning techniques to search for a structural pattern in the set of semiochemicals and then perform their classification. The pattern was established from the high belonging of a subset of these metabolites to the groups that were obtained by a grouping method based on fuzzy C-means logic; the discovered pattern corresponds to the biosynthetic pathway by which they are obtained biologically. This first classification was corroborated with Kohonen's self-organizing maps. To classify those semiochemicals whose belonging to a biosynthetic pathway was not clearly defined, we built two models of Multilayer Perceptrons which had an acceptable performance. [ABSTRACT FROM AUTHOR]

Performance tests are crucial to measure the quality of software developments, since they allow identifying aspects to be improved in order to achieve customer satisfaction. The objective of this research was to identify the optimal Machine Learning technique to predict whether or not a software development meets the customer's acceptance criteria. A dataset with information obtained from web services performance tests and the F1-score quality metric were used. This paper concludes that, although the Random Forest technique obtained the best score, it is not correct to state that it is the best Machine Learning technique; the quantity and quality of the data used in the training play a very important role, as well as an adequate processing of the information. [ABSTRACT FROM AUTHOR]

This paper presents a document classification service that allows multiple client (multi-tenant) document management systems to provide greater confidence and credibility regarding the document types assigned to documents uploaded by users. The research was carried out through the phases of CRISP-DM, where two document representation models were evaluated (bags of words with cumulative n-grams and BERT, which was recently proposed by Google) and five machine learning techniques (multilayer perceptron, random forests, k-nearest neighbors, decision trees, and naïve bayes). The experiments were carried out with data from two organizations, and the best results were obtained by multilayer perceptron, random forests, and k-nearest neighbors, which showed very similar results regarding general accuracy and recall by class. The results are not conclusive with respect to the ability to offer the service to multiple clients with a single model, since this also depends on their documents and document types. Therefore, a service is offered which is based on a microservices architecture that allows each organization to create its own model, monitor its performance in production, and update it when performance is not adequate. [ABSTRACT FROM AUTHOR]

Changes in land use and land cover lead to environmental consequences of various kinds. Digital classification of remotely sensed images is a powerful tool to assess the degree of environmental transformation due to anthropogenic factors. The aim of this work is to develop a working scheme based on remote sensing techniques and digital image classification to obtain cartographies of land use and land cover in the Segura Hydrographic Demarcation, for the period between 1986 and 2019. For this purpose, the Random Forest supervised classification algorithm has been used and the spectral bands of Landsat 5, 7 and 8 images, and four normalised indices derived from these, have been used as predictor variables. Although the overall accuracy achieved indicates that there is room for improvement in the model fit, the classifications obtained are considered reliable. With regard to the land uses and land covers obtained after the classification process, a decrease in forest use and an increase in agricultural uses, area covered by scrubland, bare soil and greenhouses can be observed. [ABSTRACT FROM AUTHOR]

Shrimp sector in Ecuador is nowadays one of the fastest-growing non-oil sectors towards the international market. In despite of this growth, to our knowledge most of the little producers of shrimps in Ecuador take important operational decisions based upon empirical knowledge, without considering historical data nor any scientific tool. In this work we implement and compare state-of-the-art statistical learning techniques for the prediction of shrimp harvest (in pounds) for a little shrimp farm located in Tenguel, Guayaquil-Ecuador. For this study we used historical information collected by the farm biologist. The data was organized and put into a digital format by the authors. Data from n=35 past harvests, corresponding to 7 cycles of production, were used to train the models. We then made predictions of shrimp harvest for the next two production cycles. We compare Multiple Linear Regression by means of ordinary least squares, CART Regression Tree, Random Forests, Multivariate Adaptive Regression Splines (MARS) and Support Vector Machines (SVM). In our analysis, MARS with no interaction terms allowed, Linear Regression with best subset variable selection and SVM with linear Kernel gave the lowest prediction error estimate by Cross Validation. Their good predictive performance was confirmed with good predictions on the next two production cycles. The use of statistical techniques can be of great help to improve predictions and therefore operational processes of small shrimp farms. [ABSTRACT FROM AUTHOR]

The objective was to estimate the forest volume (m³ ha-1) using information obtained from the Forest Planning System (SiPlaFor) and spectral satellite images data from the Landsat 8 OLI sensor in the Ejido La Victoria Pueblo Nuevo, Durango, Mexico. The non-parametric technique Random Forest was used for the forest volume estimate. The results showed that using a model with 400 decision trees, the Root Mean Square Error (RMSE) remains stable. The most important predictive variable was the Specific Leaf Area Vegetation Index (SLAVI), with an RMSE of 12.99%. The Random Forest model presented a coefficient of determination (R²) of 0.84 and RMSE of 28.8 m³ ha-1. Spectral information from a medium resolution sensor in combination with field data is a viable alternative for estimating stand volume in mixed temperate forests. [ABSTRACT FROM AUTHOR]

The purpose of this article is to make a brief introduction to five advanced machine learning prediction methods which may be useful for the development of prospective studies: logistic regression, support vector machines, gradient powered machines, random forests and neural networks. In addition, it is explained what methodology can be carried out to ensure robustness and validate these prediction models. As an example, it is presented how the use of these methods allowed to identify the most important financial variables to predict the development of innovation activities in Colombian SMEs. The results of the use of these methods may allow generating short and medium-term forecasts that serve to facilitate prospective studies with broader methods, such as the construction of scenarios, with the purpose of generating evidence-based proposals as a roadmap for long-term planning and public policy. [ABSTRACT FROM AUTHOR]

Several authors have established a large group of variables to characterise linguistic mitigation using a multidimensional database as a framework (Briz/Albelda; Albelda/others). In this study, 982 mitigation sequences have been selected from 18 spoken Spanish interviews; all of them have been identified as instances of mitigation by at least four different informants. Finally, these data have been analysed by means of three statistical methods: multiple correspondence analysis, decision trees and Random Forest. The results suggest that the variables that better characterise the three functions of mitigation (self-protection, prevention and repair) are (a) image of the speaker(s), (b) the previous speech act, (c) age, (d) syntactic position and (e) the prosodic-paralinguistic component. [ABSTRACT FROM AUTHOR]